Musk Is Building With Intel What TSMC and Samsung Feared the MOST

Kind: captions Language: en There is a rural county about an hour northwest of Houston with roughly 30,000 residents and it is about to host the largest building ever constructed by humans. Not the largest factory. The largest building, more than 100 million square feet under one roof, which is more than five times the floor area of the current record holder, the New Century Global Center in Chengdu. And inside that building, the plan is to make microchips. Logic, memory, and advanced packaging all in one place at a scale no company on Earth has ever attempted. The number attached to it is 1 terawatt of compute per year. That is roughly double what the entire United States currently produces in chips annually from a single project by a company that has never manufactured a semiconductor in its life. And that alone would be a fascinating story, but that is not the part that should worry TSMC and Samsung. The part that should worry them happened on April 7th when Intel, the company everyone in this industry had written off two years ago, signed on as the manufacturing partner. Because Intel does not bring money. Musk has money. Intel brings the one thing that cannot be bought at any price, a working recipe at the leading edge. Subscribe to the channel and let me show you exactly what this partnership changes, why 80% of this factory's output is not meant for Earth at all, and where the whole thing could still collapse. Let's start with what was actually announced because the story moved three times in five months. And most coverage is still running on the first version. On March 21st, 2026 at the decommissioned Seaholm Power Plant in Austin, Musk unveiled the project and called it the most epic chip building effort ever. The headline number was 20 to 25 billion dollars and the site was supposed to sit next to Giga Texas. Two weeks later on April 7th, Intel posted that it was joining saying it would help refactor silicon fab technology and contribute its ability to design, fabricate, and package ultra-high performance chips at scale. Intel's CEO Lip Bu Tan called it a step change in how silicon logic, memory, and packaging will get built in the future. Then in August, the real shape appeared. SpaceX and Tesla confirmed the site is not in Austin at all. It's in Grimes County on the old Gibbons Creek power plant land. First phase, $16.8 billion, 3,000 jobs. Texas extended a $30 million enterprise fund grant. Two school districts, Anderson-Shiro and Iola, approved tax abatements in mid-July. And Grimes County agreed to waive property taxes in exchange for $20 million in annual payments. The Austin site still exists, but it has been demoted to a smaller research and design fab on the Giga Texas North Campus. Now, hold that $16.8 billion figure because it is doing something strange. A Grimes County public notice tied to a June tax hearing put the initial phase at roughly $55 billion with total investment across all four planned phases reaching as high as $119 billion. Neither company has explained the gap. And Bernstein's own analysis of the full terawatt target lands somewhere around $5 trillion. So, depending on which document you read, this project costs 17 billion or 119 billion or 5,000 billion. That is not a rounding error. That is the difference between a large fab and a national infrastructure program, and we will come back to why that gap matters more than anything else in the announcement. But first, the obvious question, why would a car company and a rocket company decide to become a chip company? Because they hit a wall and Musk said so in almost exactly those words. In late January 2026, he described the situation as a choice, hit the chip wall or make a fab. And that wall is real. Look at what happened to memory over the past 12 months. DRAM spot prices have surged nearly 700% year-over-year. Global DRAM revenue jumped roughly 80% sequentially in the first quarter of 2026 alone, closing in on 100 billion dollars in a single quarter in a market that used to be famous for boom-bust cycles that left manufacturers operating at a loss. High bandwidth memory is sold out. Data centers now absorb something like 70% of all memory chips produced worldwide. Nvidia locked up roughly half a trillion dollars of future memory supply from SK Hynix in a single agreement. And SK Hynix's own CEO warned in July that this shortage probably persists well beyond 2030. So, the three companies that control over 95% of global DRAM are converting lines to HBM as fast as they can, because that is where the margin is. Every HBM wafer displaces two or more conventional DRAM wafers, which means the memory that goes into cars, into robots, into the boring parts of an autonomous driving stack is exactly the category with the least incentive to be produced. Tesla is not being outbid by other car makers, it is being outbid by hyperscalers buying for AI training clusters. Musk's framing at the announcement was diplomatic. He said they are grateful to Samsung, TSMC, Micron, and the others, and would buy all the chips they can make. But there is a maximum rate at which those companies are comfortable expanding, and that rate is much less than he would like. So, either they build the terafab or they don't have the chips. That is the honest version of the argument, and it is stronger than it sounds, because the alternative, waiting in line, is not free. Advanced node capacity is effectively spoken for years in advance. Every wafer is a fight with Apple, Nvidia, AMD, Broadcom, and anyone else willing to pre-pay billions to secure allocation. And the moment your product is compute, waiting in line means shipping late, which brings us to the part that actually changes the odds. Intel. Two years ago, Intel was the cautionary tale of this industry. It had lost ground to TSMC and AMD across nearly every product category. Its manufacturing roadmap had slipped repeatedly, and its foundry business, the attempt to make chips for outside customers the way TSMC does, had attracted almost nothing beyond government-supported work. Lip Bu Tan took over in March 2025 and said something remarkable for a CEO in his position. Intel would not commit to building factories on its most advanced process unless it had committed external customers. That process is called 14A, and it is the single most important technical detail in this entire story, because Terafab is not being built on 2 nanometers anymore. On Tesla's Q1 2026 earnings call, Musk confirmed the plan is Intel 14A. He called it state-of-the-art and admitted openly that it is not totally complete yet, arguing that by the time Terafab scales up, 14A will be mature enough for prime time. So, what is 14A, and why does it matter so much? Three things stack inside it. The first is Ribbon FET 2, Intel's second generation of gate-all-around transistors. Gate-all-around is the architecture that replaced FinFET at the leading edge. In FinFET, you lift the channel up and wrap the gate around three sides. In gate all around, you stack horizontal nanosheets, each only a few nanometers thick, and wrap the gate completely around every single one. Better electrostatic control, less leakage, and dramatically more ways for the process to go wrong because now every layer thickness, every spacing, every interspacer has to be controlled with atomic precision across billions of devices. The second is Power Direct, and this one is genuinely ahead of the field. Intel's 18A introduced PowerVia, the industry's first production backside power delivery network, where the power rails run underneath the transistor layer instead of competing for space with signal wiring on top. Power Direct goes a step further and delivers power contacts directly to each transistor source and drain through dedicated backside contacts rather than routing through through-silicon vias. When 14A reaches high-volume manufacturing, Intel will be two full generations of backside power ahead of TSMC's production nodes. TSMC's first backside powered node, a 16, is only targeted for production in late 2026. And the third is high-NA EUV, which is where this gets really interesting. Standard EUV lithography prints with a numerical aperture of 0.33. High-NA pushes that to 0.55, which buys you finer resolution and, more importantly, lets you print in fewer exposures. Intel installed ASML's Twinscan EXE:5200B, the first commercial high-NA tool, and passed acceptance testing. Each of those machines cost around $380 million versus roughly $235 million for the previous generation. Expensive. But Intel has reported cutting the process steps required for certain layers from around 40 to fewer than 10. Fewer steps means shorter cycle time, fewer masks, fewer chances to introduce a defect. Here's the distinction that matters. On 18A, high NA was inserted selectively on certain layers after the node's design was largely fixed. 14A was designed around high NA from inception. It is the first process in the world built that way. And ASML confirmed in July 2026 that Intel became the first company to ship logic chips made with high NA EUV. TSMC has said it will not use high NA at A14. It's rough equivalent node, also targeted for volume production in 2028. So, for the first time in about a decade, there is a leading-edge process where Intel is not chasing. It is ahead on backside power, ahead on high NA insertion, and according to CFO David Zinsner, speaking at Deutsche Bank's technology conference in 2026, 14A is showing better defect density reduction than any Intel node since 22 nanometers. Morgan Stanley pegged early 14A defect density around 0.5, which for a pre-production node is respectable. Intel says 14A maturity, yield, and performance are outpacing 18A at the same point in its life. And that is what Musk actually bought. Not machines, machines you can order, recipes you cannot. Because this is the part that people outside the industry consistently underestimate. A modern fab runs thousands of process steps, and every one of them interacts with the others. Lithography affects etch, etch affects deposition, deposition affects electrical behavior. Every step adds variation, and variation is what kills yield. TSMC did not become TSMC by buying better equipment than everyone else. They bought the same equipment and then spent decades and hundreds of billions learning what to do with it. That learning is not written down anywhere you can purchase. Patrick Moorhead put it bluntly after the August announcement. You can't build a 2-nanometer fab without a production-ready process flow from someone. That sentence is the whole reason Intel is in this project. Intel is the someone. Now, let's talk about why TSMC and Samsung read this announcement very differently from each other. For Samsung, this is personal. In July 2025, Samsung landed a contract widely reported at around $16.5 billion to build Tesla's AI 5 and AI 6 chips. And that contract was the anchor tenant that let Samsung finish its Taylor, Texas fab and move equipment in. Tesla was the customer that made Taylor make sense. By July 2026, Samsung was taping out AI 5 on its 2-nanometer gate-all-around process at Taylor, a node that had been assumed to be reserved for the later AI 6 chip, which suggests Samsung's 2-nanometer yields finally crossed the threshold where the process becomes viable for high volume. Musk has publicly said Samsung's Texas fab outclasses TSMC's US fabs on tooling. And when Tesla approached Samsung about Terafab itself, Samsung reportedly declined to participate directly and instead offered to allocate additional capacity at Taylor, which is a completely rational answer and also possibly the wrong one. Because Samsung is now in the position of ramping a multi-billion-dollar American fab around a customer that has publicly announced its intention to build its own supply. In the near term, Samsung probably wins. Tesla needs bridge capacity while Terafab is under construction, and that means more orders, not fewer. In the long term, Samsung risks losing the anchor customer for its most advanced American node. For TSMC, the threat is different and slower. TSMC holds around 70% of the foundry market and is not going to lose that to a company that has never made a wafer. Terafab is not a near-term technological challenge to TSMC's logic business. The pressure point is packaging. TSMC's chip on wafer on substrate lines, CoWoS, are the bottleneck through which basically every high-end AI accelerator on the planet passes. They are capacity constrained, they are backlogged, and they are concentrated in Taiwan. Intel brings EMIB, its embedded multi-die interconnect bridge, and Foveros 3D stacking, both manufactured on US soil. Both capable of integrating logic and memory chiplets. John Lorenz at Yole Group described the logic exactly. Hyperscalers need a second source. Intel's EMIB on American soil can fill that gap, and Terafab serves as a high-profile proof of execution. That is the actual strategic move here. Terafab may or may not ever hit a terawatt, but if it gives Intel a credible high-volume American alternative to CoWoS, it changes the geometry of the entire AI supply chain, and it does that years before the first Terafab wafer exists. Stephen Ezell at the Information Technology and Innovation Foundation called the announcement a very significant win for Intel as it builds a foundry business alongside its core chip-making and said it will bolster Intel's ability to manufacture chips for AI and mobility applications. Adrian Sanchez at Yole Group went further. TerraFab is first and foremost a breakthrough for Intel Foundry, adding Tesla, SpaceX, and xAI as anchor customers, and giving the Foundry the credibility it has been chasing for years. But Sanchez also said, "The thing that reframes this entire project, and it is the first big catch, advanced node fabrication requires EUV lithography, and only ASML builds those machines. ASML shipped 48 EUV systems in 2025. Its order book is fully allocated to TSMC, Samsung, and Intel through 2027. TerraFab has no confirmed ASML orders. As a brand new market entrant with no history, no relationship, and no slot in the queue, it faces exactly the procurement problem you would expect." So Sanchez's conclusion is this, "Rather than building a greenfield fab from scratch and waiting years for equipment, TerraFab can leverage Intel Foundry's existing 14A process technology, equipment allocations, and packaging infrastructure. In other words, TerraFab is likely more an Intel fab expansion with Tesla, SpaceX, and xAI as anchor customers than a standalone manufacturing venture." Read that again because it dissolves most of the skepticism and creates a new problem at the same time. If TerraFab is effectively Intel capacity wearing a Musk badge, then the equipment question is solved, the process question is solved, and the timeline gets dramatically more believable. But it also means that Terrawatt is not really being built by a new entrant reinventing the fab. It is being built by Intel at Intel's pace, constrained by Intel's own allocation from ASML, the same allocation Intel needs for its own products. And lithography is only one tool class. A modern fab needs deposition, etch, ion implantation, metrology, inspection, cleaning, polishing, and packaging. Hundreds of tools, most with 12 to 24-month lead times, then months to install, months to calibrate, and years before anyone actually masters them. Scaling from a pilot line to the stated long-term target of 1 million wafer starts per month is not one factory. It is dozens of factory equivalents sequenced. There is one detail in the plan that shows how seriously somebody is thinking about the physical problem, though. To move material across a campus this large, The Boring Company is building tunnels underneath the site. The stated reason is not just automation, it is vibration. EUV scanners align patterns to a fraction of a nanometer, and they are exquisitely sensitive to mechanical noise, which is why fabs are built on massive isolated slabs, and why site selection normally weighs seismic stability first. Running your internal logistics underground instead of on the surface removes a whole category of disturbance from the most expensive tools in the building. That is not a marketing detail. That is somebody who has read the tool specifications, which brings us to the second catch, and this one nobody at the announcement wanted to discuss. Memory. The Terafab pitch is logic, memory, and advanced packaging under one roof. That word memory is doing enormous work. John Lorenz at Yole Group ran the numbers. Filling Terafab's eventual demand would require something like 10 million wafers per month just for memory. That is five times the current global production of DRAM. Five times. Not 5% more. Five times the output of an entire industry that took 40 years and hundreds of billions of dollars to build. And memory is not logic with different masks. It is a fundamentally different factory, different process flows, different tool sets, different yield economics, far more dies per wafer, and a completely different manufacturing culture. Lorenz was direct about it. Terafab is years from producing anything. Memory manufacturing requires entirely separate expertise that neither Tesla, SpaceX, nor Intel possesses. And he doubts Terafab could produce its own memory without a partnership from the incumbents. Intel, remember, exited the memory business decades ago. That is not a gap you close with capital. It is a gap you close by hiring the people who already know from companies that have every reason not to let them go. And there is a second-order risk here that gets almost no attention. Sanchez pointed out that the real pricing danger is not Musk building his own DRAM. It is the eventual capacity overshoot when all these new fabs come online around 2028 to 2029. The memory industry's entire history is a cycle of shortage, panic expansion, glut, and collapse. Right now, the panic expansion phase is running at full speed across Samsung, SK Hynix, Micron, and now this. Everyone is building for the demand curve of 2026. Nobody has ever correctly predicted the demand curve of 2029. Now, the third catch, and this one is physical. Power. A leading-edge fab at the announced scale draws somewhere between half a gigawatt and 10 gigawatts continuously, depending on the phase. And Texas has a problem. The ERCOT interconnection queue for large loads has reached roughly 474 gigawatts of requests. About 90% of them data centers against an all-time peak demand for the entire grid of under 100 gigawatts. Governor Abbott ordered ERCOT and the Public Utility Commission to pause pending data center interconnections until each project passes an audit on water, power, incentives, and ownership. So, Terafactory is not going to wait in that queue. According to the executed tax abatement agreement with Grimes County, the facilities are expected to be powered by on-site power plants and are not expected to use electricity from the grid. SpaceX confirmed the plan, gigawatt-scale natural gas generation plus very large battery arrays. At the August 6th public meeting, SpaceX's representative told Grimes County residents the company would be bringing its own power so that Terafactory's demand does not raise costs for other Texas rate payers. And there is a technical argument for gas that has nothing to do with emissions. Fabs are brutally sensitive to power quality. A voltage sag of a fraction of a second can scrap every wafer in process across the entire line. Gas turbines paired with batteries are easier to condition at gigawatt scale than solar plus storage, and they can be operational in months while utility-scale solar interconnection typically runs 3 to 5 years under current queue conditions. When your published target is first chip production in late 2027, months versus years decides the architecture. Still, the irony is hard to miss. Tesla is a partner on this project. Tesla sells solar, Powerwall, and Megapack. Tesla and Sunrun launched a 16-gigawatt virtual power plant aimed directly at data centers in June. xAI itself bought $269 million of Tesla Megapacks earlier this year. And the answer for Terafactory is gas from APR Energy. The gas turbine company Musk personally bought for around a billion dollars in July on top of more than $2.8 billion dollars has separately committed to turbines. The vision is solar. The buildout is gas. Location does solve some problems, though. Helium is used as a cooling and process gas in advanced lithography and packaging, and Asian foundries lean heavily on imports from the Gulf region, a supply line that had production outages in early 2026. The United States produces close to half the world's helium from reserves in Texas, Wyoming, Kansas, and Oklahoma. So, Terafab starts with a genuine raw material advantage that TSMC in Taiwan simply does not have. But, as Lorenz cautioned, bare silicon wafers still mostly come from Asian suppliers, and the water and power needs of semiconductor manufacturing at this scale are not trivial. You can build the building in Texas. You cannot yet source the entire input stack there. Now, the loop I opened at the start. Because everything so far assumes this is a factory for AI data centers and cars. It mostly isn't. Roughly 20% of the planned output goes to the AI 5 and AI 6 architectures, edge inference chips for Tesla's autonomy stack, the cyber cab, and the Optimus humanoid program. The other 80% is designated for space. The chip is called D3, a radiation-hardened processor built on the Dojo 3 architecture, and it is meant for SpaceX's orbital data center satellites, internally designated AI sat mini. Each satellite is described as generating around 100 kilowatts of power with solar arrays exceeding 170 meters in length. SpaceX has filed an FCC application covering up to 1 million such satellites, although President Gwynne Shotwell has indicated that number is an intentional ceiling rather than a firm plan. The thesis behind orbital compute is that space gives you continuous solar power and a vacuum for thermal rejection without competing for terrestrial grid capacity or water. And silicon behaves very differently up there. Space does not fail your hardware all at once. It degrades it. High energy particles flip bits, corrupt calculations, and accumulate damage in transistors over time. Standard commercial logic does not survive that environment, which is why radiation hardened parts exist and why they are so expensive. Fewer dies survive per wafer. Every part goes through qualification far beyond normal production testing. And in the United States, rad hard silicon sits close to defense technology. So approvals add time on top of everything else. There is also a genuine engineering tension buried in this plan. Radiation tolerance and aggressive scaling pull in opposite directions. The smaller and more tightly packed the transistor, the less charge it takes to disturb it and the more likely a single particle strike takes down a logic block, which means the space chip and the AI inference chip do not want the same process, do not want the same design rules, and in a shared facility, do not necessarily want to share tools. A vertically integrated megafab that runs bleeding edge logic, memory, advanced packaging, and radiation hardened silicon under one roof is not one hard problem. It is four hard problems that contaminate each other's assumptions. But if it works, the payoff is exactly the kind of thing vertical integration is good at. Rad hard chips today are priced like aerospace components because they are made in tiny volumes with enormous qualification overhead. Change the volume, own the process, and the cost curve moves. That is the same argument SpaceX already won once with reusable rockets. Where cutting launch cost by roughly an order of magnitude did not just make the existing market cheaper. It created markets that did not exist. Now, the money. Because the second loop I opened was that gap between $16.8 billion and $5 trillion, and this is where it gets resolved or doesn't. On June 12th, 2026, SpaceX went public on the Nasdaq under the ticker SPCX, pricing 555.6 million shares at $135 each and raising about $75 billion. That is the largest initial public offering in history at a valuation near $1.75 trillion. Roughly 30% of the offering was directed to retail investors, which is extraordinary compared to the single-digit allocations typical of large listings. Shares surged on debut and briefly pushed the market cap above $2 trillion, peaking at $225.64 intraday on June 16th. By late August, the stock was trading around $137, barely above the IPO price. That round trip matters because the financing logic of this entire project runs through the public market. And SpaceX's first quarterly report as a public company showed a $541 million loss on $7.8 billion in revenue. In the first quarter of 2026 alone, the XAI unit burned $7.72 billion and posted a $2.47 billion operating loss. This is not a company throwing off spare cash. It is a company that just raised the largest sum in IPO history and still has an AI division consuming billions a quarter. Meanwhile, Tesla's CFO acknowledged that the full cost of Terafab is not incorporated into Tesla's 2026 capital expenditure plan. And Intel, for its part, has disclosed no contract value for Terafab in its public filings. No dollar commitment, no equity stake, no capacity allocation. Intel's role in this partnership is in public terms narrow and unpaid and strategically enormous. Intel raised roughly $20 billion in capital in the same period. And the US government took a stake in the company in August 2025 holding around 8 to 10% of shares outstanding. Intel's cash and short-term investments stood at $29.7 billion as of June 2026. As of late June 2026 against $48.5 billion in total debt. So the financing picture is a public market that has already cooled on the story, a partner that hasn't committed public dollars, an anchor company burning cash on AI, and a project whose own filings disagree with each other by a factor of seven. That is why the tax structure is worth looking at. Terafab filed eight tax incentive applications with the two local school districts asking to freeze school maintenance and operations taxes at roughly 48% of appraised value for 10 years per phase and asking the state to cover the difference through the jobs, energy, technology, and innovation program. The stated justification was the cost differential between Texas and competing jurisdictions. Grimes County waived property taxes in exchange for $20 million a year. Texas added the $30 million enterprise fund grant. None of that is unusual for a project of this size. Samsung got similar treatment for Taylor. But it tells you something about how the numbers close. At $16.8 billion phase one is financeable. At $119 billion across four phases it needs the IPO proceeds and then some. At $5 trillion it needs something that does not currently exist in private capital markets. So does this work? Here's the strongest version of the case that it does. The demand is real and documented. The memory shortage alone is forcing decisions at companies far larger than Tesla. The site is secured, the incentives are signed, the power plant bypass is the single worst bottleneck in Texas, and the research fab in Austin is already under construction. Most importantly, the largest execution risk, not knowing how to run a leading edge process, has been outsourced to the one American company that does. Intel gets the anchor customer that justifies 14A. Musk gets a recipe he could not develop in a decade, and both of them get a domestic packaging alternative to a Taiwanese bottleneck that every hyperscaler on Earth is currently worried about. If you strip away the terawatt headline, what remains is a large, well-sited, well-financed American fab expansion with a very motivated first customer. That version is not crazy at all. And here is the strongest version of the case that it doesn't. The terawatt target requires roughly 50 times the compute output of the current global semiconductor industry, and no public path to it has been published. The equipment queue is allocated through 2027, and Terafab has no confirmed orders of its own. Memory at the stated scale would require five times the world's DRAM output and expertise that none of the three named partners has. The compute demand forecast assumes Optimus, Cybercab, and an orbital constellation all hit their projected scale. And [snorts] every one of those programs has slipped before. Tesla dissolved its Dojo D1 team in 2025 and pivoted to third-party silicon, which is not the track record of a company that finishes chip mega projects. And Musk's own published target of first chip production in late 2027 is, for a fab of this scale and complexity, extraordinarily aggressive. There is also the risk that nobody talks about until it arrives, which is utilization. A fab only makes money when it is full. Depreciation does not care about demand. Intel learned that lesson the hard way. It led the industry, slipped on a single process node, and then carried the full weight of its own manufacturing while Nvidia and Apple stayed fabless, flexible, and fast. Extreme vertical integration gives you control, and it takes away your safety net. The companies now betting on it are betting that compute demand never softens. What Terafab really represents, regardless of whether it ever reaches a terawatt, is the moment AI compute stopped being treated as a market problem and started being treated as national infrastructure. Hyperscalers are building their own gas plants. A rocket company is building the largest structure on Earth to make its own silicon. And a chip maker that was nearly written off is suddenly the most strategically important manufacturing asset in the country. As of today, Terafab has a site, signed incentive agreements, a funded first phase, 3,000 committed jobs, an Intel partnership, and a research fab under construction. That is substantially more than a keynote slide. What it does not have is a groundbreaking date for the main campus, a first wafer date, a published path to the terawatt, or a single confirmed EUV tool of its own. So, here is the question I would put to you. Is this the most aggressive and correct vertical integration bet in modern semiconductor history, the moment America stopped renting its most critical technology, or is it a nine-figure marketing campaign for an Intel fab expansion that would have happened anyway, wrapped in a number nobody expects to hit? Tell me in the comments which side you land on and what would change your mind. If this breakdown was useful, share it with someone who still thinks a chip factory is just a very expensive building, and [snorts] subscribe so you don't miss what happens when the first wafers actually run.